WO2018107492A1 - 基于直觉模糊随机森林的目标跟踪方法及装置 - Google Patents

基于直觉模糊随机森林的目标跟踪方法及装置 Download PDF

Info

Publication number
WO2018107492A1
WO2018107492A1 PCT/CN2016/110478 CN2016110478W WO2018107492A1 WO 2018107492 A1 WO2018107492 A1 WO 2018107492A1 CN 2016110478 W CN2016110478 W CN 2016110478W WO 2018107492 A1 WO2018107492 A1 WO 2018107492A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
result
intuitionistic fuzzy
node
sample
Prior art date
Application number
PCT/CN2016/110478
Other languages
English (en)
French (fr)
Inventor
李良群
李俊
谢维信
刘宗香
李小香
Original Assignee
深圳大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Priority to PCT/CN2016/110478 priority Critical patent/WO2018107492A1/zh
Publication of WO2018107492A1 publication Critical patent/WO2018107492A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion

Definitions

  • the invention relates to the field of target tracking, in particular to a target tracking method and device based on intuitionistic fuzzy random forest.
  • Online target tracking is a hot research topic in computer vision. It is of great significance for high-level visual research such as motion recognition, behavior analysis and scene understanding, and has wide applications in video surveillance, intelligent robots, human-computer interaction and other fields. prospect.
  • the missed detection target cannot find the detected observation object associated with it, and it is impossible to find valid information for the track update of these missed detection targets through the data association, and the track accuracy is lowered.
  • the present invention proposes a target tracking method based on an intuitionistic fuzzy random forest.
  • the object tracking method based on the intuitionistic fuzzy random forest includes: performing motion detection on the current video frame, detecting the possible moving object as an observation result; and correlating the observation result with the prediction result of the target, wherein the prediction result is at least using the previous video.
  • the target includes a reliable target and a temporary target; and trajectory management is performed on the unrelated observation and prediction results, including online tracking acquisition candidates for the prediction results of the unassociated reliable target
  • the candidate results are matched by the intuitionistic fuzzy random forest of the unassociated reliable target
  • the trajectory of the target of the current frame is obtained by using the association result and the matching result, including the candidate result of the successful target matching the successful target with successful matching
  • the prediction result is filtered and updated to obtain the trajectory; the trajectory of the target of the current frame is used for prediction, and the intuitionistic fuzzy random forest is updated for the reliable target with successful association or matching success.
  • the object tracking device based on the intuitionistic fuzzy random forest comprises: a processor and a camera, the processor is connected to the camera; and the processor is configured to perform motion check on the current video frame acquired from the camera Measuring, detecting the possible moving object as an observation result; correlating the observation result with the prediction result of the target, wherein the prediction result is obtained by predicting at least the trajectory of the target of the previous video frame, and the target includes the reliable target and the temporary target Trajectory management of unrelated observations and prediction results, including online tracking of candidate results of unassociated reliable targets, and candidate results by using intuitionistic fuzzy random forests of unassociated reliable targets Matching; using the correlation result and the matching result to obtain the trajectory of the target of the current frame, including using the candidate result of the matching success to filter and update the prediction result to obtain the trajectory; and using the trajectory of the target of the current frame Predict and update the intuitionistic fuzzy random forest for reliable targets that are successful or
  • the beneficial effects of the present invention are: obtaining candidate results by performing online tracking on the prediction results of unassociated reliable targets, and matching the candidate results with the intuitionistic fuzzy random forests of the unassociated reliable targets, and if the matching is successful, using the matching
  • the successful candidate results filter and update the prediction result of the reliable target to obtain its trajectory, so that in the event that the missed detection occurs and the target cannot find the associated observation object, the intuitionistic fuzzy random forest can be used to find the matching and can be used for matching.
  • the trajectory filters the updated candidate results, thereby improving the accuracy of the target trajectory and improving the performance of the target tracking.
  • FIG. 1 is a flow chart of a first embodiment of a target tracking method based on an intuitionistic fuzzy random forest according to the present invention
  • FIG. 2 is a schematic diagram of a hard decision function and a fuzzy decision function of a branch node in an example of a second embodiment of the object tracking method based on the intuitionistic fuzzy random forest;
  • FIG. 3 is a schematic diagram of a fuzzy decision function and an intuitionistic fuzzy decision function of a branch node in an example of a second embodiment of the object tracking method based on the intuitionistic fuzzy random forest;
  • FIG. 4 is a flow chart of a third embodiment of a target tracking method based on intuitionistic fuzzy random forest according to the present invention.
  • FIG. 5 is a flow chart of a fourth embodiment of a target tracking method based on intuitionistic fuzzy random forest according to the present invention.
  • FIG. 6 is a flowchart of the feature selection criterion training in the fourth embodiment of the target tracking method based on the intuitionistic fuzzy random forest of the present invention
  • FIG. 7 is a flow chart of a fifth embodiment of a target tracking method based on an intuitionistic fuzzy random forest according to the present invention.
  • FIG. 8 is a flowchart of a sixth embodiment of a target tracking method based on an intuitionistic fuzzy random forest according to the present invention.
  • FIG. 9 is a flowchart of a seventh embodiment of a target tracking method based on an intuitionistic fuzzy random forest according to the present invention.
  • FIG. 10 is a schematic structural diagram of a first embodiment of a target tracking apparatus based on an intuitionistic fuzzy random forest according to the present invention.
  • the first embodiment of the target tracking method based on the intuitionistic fuzzy random forest of the present invention includes:
  • the motion detection algorithm such as frame difference method, optical flow method and background subtraction method is used to detect the motion of the current video frame, so as to find out the pixels belonging to the foreground of the motion, supplemented by median filtering and simple morphological processing, and finally obtain the current video.
  • Possible moving objects in the frame are used as observation objects.
  • An observation object is an image block in the current video frame. Generally, the shape of the observation object is a rectangle.
  • Targets include reliable targets for stable tracking and temporary targets for unstable tracking.
  • the target state in this step that is, whether each target is marked as a reliable target or a temporary target, is determined by the trajectory management of the previous video frame.
  • the temporary target includes a new target established by the observation that the previous video frame is a candidate result that is not associated and is not a successful match, and a target whose consecutively associated successful number of frames is less than or equal to the first frame number threshold and has not been deleted.
  • a reliable target includes a target whose number of consecutively successful frames is greater than the first frame number threshold and has not been deleted.
  • the prediction result of the target is obtained by predicting at least the trajectory of the target of the previous video frame.
  • S3 Perform trajectory management on uncorrelated observations and prediction results, including performing on-line tracking of candidate results for unassociated reliable targets, and using intuitionistic fuzzy random forest pairs of unassociated reliable targets Candidate results are matched.
  • a plurality of image blocks are selected as candidate results in the prediction result position of the reliable target and the surrounding range, and the size of the image block is generally consistent with the size of the prediction result, and the size of the specified range and the number of candidate results are generally experienced.
  • the value is determined.
  • Candidate results can include unrelated observations within a specified range. Adjacent candidate results may not overlap each other or may partially overlap.
  • Intuitionistic fuzzy random forests with unassociated reliable targets are used as classifiers, and the classification results of the classifiers are reliable and non-reliable.
  • the candidate result is calculated as the intuitionistic fuzzy membership of the test sample belonging to the reliable target category.
  • the matching is successful, 0.5 ⁇ 1 ⁇ 1 and 0.5 ⁇ 2 ⁇ 1.
  • the candidate target with the largest intuitionistic fuzzy membership degree can also be selected from the candidate results, and then Determining whether the intuitionistic fuzzy membership degree of the selected candidate result is greater than the first threshold value ⁇ 1 , and whether the appearance feature similarity measure of the prediction result is greater than the second threshold value ⁇ 2 , and if the foregoing two conditions are met, the matching is successful.
  • the target status is updated based on the association result and the matching result, including the establishment, deletion, and state modification of the target.
  • the method includes: establishing a new temporary target for the observation result of the candidate result that is not associated and not successfully matched; changing the temporary target whose continuous association is greater than the first frame number threshold ⁇ 1 into a reliable target; deleting the continuous association
  • the number of successful frames is greater than the temporary target of the second frame number threshold ⁇ 2 ; the number of frames in which the consecutive association is unsuccessful is greater than the third frame threshold ⁇ 3 , and the matching result is a reliable target of matching failure, and the matching result is that the matching failure means
  • the intuitionistic fuzzy random forest calculation candidate result is used as the test sample, and the intuitionistic fuzzy membership degree belonging to the reliable target category is less than or equal to the sixth threshold ⁇ 6 and satisfies 0 ⁇ 6 ⁇ ⁇ 1 .
  • ⁇ 1 is a positive integer greater than 1
  • ⁇ 2 and ⁇ 3 are both positive integers and satisfy ⁇ 3 ⁇
  • S4 Acquire the trajectory of the target of the current frame by using the association result and the matching result, perform prediction by using the trajectory of the target of the current frame, and update the intuitionistic fuzzy random forest for the reliable target with successful association or matching success.
  • the reliable target matching success is filtered and updated by the candidate result of the matching success to obtain the trajectory.
  • the associated successful target uses its associated observations to filter and update its prediction results to obtain the trajectory.
  • the new temporary target will be the corresponding observation result as the trajectory, and the temporary target and the association that are not successful and not deleted will not be associated.
  • a reliable target that succeeds and matches unsuccessfully and has not been deleted has its predicted result as a trajectory.
  • the trajectory of the target of the current frame is used for prediction, and the obtained result can be used as the target prediction result for the target tracking of the next frame.
  • the Kalman filter is used to predict the trajectory of the target of the current frame to obtain the prediction result of the target of the next frame, and the Kalman filter can also be used for the prediction result and the corresponding observation result/ The candidate results are filtered to obtain the trajectory of the target.
  • the target image block corresponding to the reliable target with successful association or matching success is used to update the intuitionistic fuzzy random forest.
  • the target image block may not include the trajectory information of the target, for example, the associated successful observation object or the candidate result of the successful matching.
  • the step of updating the intuitive fuzzy random forest and the step of acquiring and predicting the trajectory of the foregoing target are not executed. limit.
  • the target image block may also include trajectory information of the target, such as an image block at the location of the trajectory of the reliable target, and the step of updating the intuitive ambiguous random forest at this time should be performed after the step of acquiring the trajectory of the aforementioned target.
  • the prediction result of the unassociated reliable target is obtained by online tracking to obtain the candidate result, and the candidate fuzzy result is matched by the intuitionistic fuzzy random forest of the unassociated reliable target. If the matching is successful, the matching is successful.
  • the candidate result filters and updates the prediction result of the reliable target to obtain its trajectory, so that in the case where the missed detection occurs and the target cannot find the associated observation object, the intuitionistic fuzzy random forest can be used to find the matching and can be used for it.
  • the trajectory filter updates the candidate results, thereby improving the accuracy of the target trajectory and improving the performance of the target tracking.
  • the second embodiment of the target tracking method based on the intuitionistic fuzzy random forest is based on the first embodiment of the target tracking method based on the intuitionistic fuzzy random forest of the present invention, and the candidate result is used as the intuitionistic fuzzy of the test object belonging to the reliable target category.
  • the membership degree P(c m
  • w) is:
  • c is the category label of the test sample
  • m is the reliable target category
  • w is the test sample
  • T is the number of intuitionistic fuzzy decision trees in the intuitionistic fuzzy random forest
  • M is the category set of the training samples that generate the intuitionistic fuzzy random forest
  • ⁇ 1 (c m
  • w) is the intuitionistic fuzzy membership degree of the test sample w calculated by the t-th intuitionistic fuzzy decision tree belonging to the reliable target class m.
  • the intuitionistic fuzzy decision tree intuitions the branch node output decision so that the same sample will pass through the output left branch and the output right branch of the branch node with different intuitionistic fuzzy membership degrees, and finally reach multiple leaf nodes. Therefore, the classification result of the intuitionistic fuzzy decision tree needs to comprehensively consider the information of multiple leaf nodes.
  • ⁇ t (c m
  • w) is defined as:
  • B t is a set of all leaf nodes that the test sample w arrives
  • b is a leaf node that the test sample w arrives
  • h % (w) is the leaf node b as the current node.
  • the test sample w belongs to the intuitionistic fuzzy membership of the current node
  • the confidence level for the leaf node b to predict the class m is defined as:
  • x j is the training sample reaching the leaf node b
  • n b there are n b
  • c j is the category of the training sample x j
  • ⁇ ( ⁇ ) is the Dirac function
  • h % (x j ) is the leaf node b as the current node
  • the training sample x j belongs to the intuitionistic fuzzy membership of the current node.
  • the intuitionistic fuzzy membership degree of the training sample and the test sample belonging to the current node is calculated in the same way.
  • the sample x belongs to the current node's intuitionistic fuzzy membership degree h % (x) as the output of all the branch nodes that belong to the current node.
  • the product of the intuitionistic fuzzy membership of the path specifically defined as:
  • D is the set of all branch nodes that the sample passes before reaching the current node
  • d is a branch node in the set
  • l represents the output left branch of the branch node
  • r represents the output right branch of the branch node
  • the sample includes test samples and training samples.
  • the training samples include positive training samples and negative training samples. Positive training samples refer to training samples that are target categories, and negative training samples refer to training samples that are non-target categories. If the current node is the root node, D is empty, and h % (x) cannot be calculated using equation (4).
  • h % (x) 1
  • h % (x) 1/n 1
  • h % (x) 1/n 0 .
  • Intuitionistic fuzzy membership of the output path of the branch node d that the sample passes to reach the current node defined as:
  • the output path of the branch node d that should pass according to the sample is the left branch or the right branch, from equation (5) with Select one of the corresponding expressions in the expression.
  • h(x d ) is the intuitionistic fuzzy output decision function of the branch node d.
  • branch nodes of the traditional binary decision tree adopt hard decision, and the branch node output decision function is defined as:
  • x d is the eigenvalue of the sample x of the branch node d
  • ⁇ is the feature threshold.
  • 0 corresponds to the branch node output left branch
  • 1 corresponds branch node output right branch.
  • the conventional hard decision function represented by the equation (19) is blurred using a sigmoid function (i.e., a Sigmoid function).
  • x d is the eigenvalue of the sample x of the branch node d
  • ⁇ is the characteristic threshold
  • ⁇ is a constant parameter for controlling the degree of tilt of the Sigmoid function
  • ⁇ is the standard deviation of the eigenvalue.
  • the dotted line in the figure represents the hard decision function defined by equation (19), and its output jumps at the feature threshold; the solid line represents the fuzzy decision function defined by equation (7), whose output monotonically changes continuously according to the eigenvalue of the sample. And equal to 0.5 at the feature threshold.
  • the intuitionistic fuzzy point operator is used to further extend the fuzzy decision function based on Sigmoid function to the intuitionistic fuzzy decision function.
  • ⁇ A U ⁇ [0, 1]
  • ⁇ A (u) denotes the membership degree of the element u in the set U belonging to A
  • v A U ⁇ [0, 1]
  • v A (u) represents the element in the set U u belongs to the non-affiliation of A, and for any u:
  • the fuzzy intuitionistic index of the element u in the set U belonging to A is defined as:
  • the fuzzy intuition index ⁇ A (u) represents the uncertainty information of the element u relative to the intuitionistic fuzzy set A. If the value of ⁇ A (u) is small, it indicates that the membership value of element u belongs to A is relatively accurate; if the value of ⁇ A (u) is large, it means that the membership value of element u belongs to A has greater uncertainty. Sex. Compared with fuzzy sets, intuitionistic fuzzy sets can reflect the information of membership, non-affiliation and fuzzy intuition index, which is beneficial to better deal with the information of uncertainty.
  • Intuitionistic fuzzy point operator Transform the intuitionistic fuzzy set A into an intuitionistic fuzzy set with the following fuzzy intuitionistic exponents:
  • the fuzzy intuition exponent ⁇ A (u) is divided into: (1- ⁇ u - ⁇ u ) n ⁇ A (u), ⁇ u ⁇ A (u)(1-(1- ⁇ u - ⁇ u ) n )/ ( ⁇ u + ⁇ u ) and ⁇ u ⁇ A (u)(1-(1- ⁇ u - ⁇ u ) n )/( ⁇ u + ⁇ u ) are three parts, which respectively represent the unknown in the original uncertain information. , affiliated and non-affiliated parts.
  • Equation (30) shows that the intuitionistic fuzzy point operator
  • the fuzzy intuition index of the intuitionistic fuzzy set A can be reduced.
  • Intuitive fuzzy point operator New information can be extracted from the uncertainty information of the element u relative to the intuitionistic fuzzy set A, and the degree of utilization of the uncertain information is improved.
  • the intuitionistic fuzzy output decision function h(x d ) of the branch node d obtained by the intuitionistic fuzzy generalization is defined as:
  • k is the number of operators and is a positive integer.
  • ⁇ (z) in equation (6) is a fuzzy intuitionistic index, which is defined as:
  • ⁇ is a constant parameter, 0 ⁇ ⁇ ⁇ 1, for example 0.8.
  • x d ⁇ ⁇ , z g (x d )
  • x d ⁇ ⁇ , z 1 - g (x d ).
  • ⁇ in equation (6) is a scale factor for extracting membership information from the fuzzy intuition index
  • ⁇ is a scale factor for extracting non-subordinate information from the fuzzy intuition index
  • Equation (6) represents the intuitionistic fuzzy membership of the sample belonging to the right branch of the branch node output.
  • Intuitionistic fuzzy point operator It is able to extract new useful information from the uncertain information.
  • the graph of the branch node intuitionistic fuzzy output decision function is as shown in FIG. 3.
  • the dashed line in the figure represents the fuzzy decision function defined by equation (7); the solid line represents the intuitionistic fuzzy decision function defined by equation (6).
  • the value of the operator number k and the feature threshold value ⁇ in the equation (6) can be determined by updating the feature selection criterion in the process of the intuitionistic fuzzy decision tree, or can be determined by other methods such as empirical values.
  • the branch nodes of the traditional binary decision tree adopt hard decision.
  • the test sample can only select one of the left and right branches according to the feature attribute to reach the next layer node, and finally reach a leaf node.
  • the test sample category is determined by the arriving leaf node. The category is determined.
  • Such a hard decision decision tree is not robust to sample noise. When the sample is subjected to strong noise interference, its eigenvalue will change greatly, which may cause the branch of the sample to change and reduce the accuracy of the decision tree.
  • fuzzy decision tree In the prior art, a fuzzy decision tree is proposed.
  • the fuzzy set theory is applied to the training and reasoning process of decision trees.
  • the representation ability of fuzzy set theory is used to improve the processing ability of traditional decision trees for noisy data and incomplete data.
  • the fuzzy decision tree can process the eigenvalues with uncertainty, it needs to perform fuzzy semantic processing on the sample features, and the sample features used in the target tracking are mostly numerical features, and the feature dimension is high, which makes Fuzzy semantic processing of sample features becomes difficult.
  • the intuitionistic fuzzy random forest in this embodiment comprehensively considers the intuitionistic fuzzy membership degree calculated by each intuitionistic fuzzy decision tree to obtain the classification result of the test sample, and the classification performance of the intuitionistic fuzzy random forest is better than that of the single intuitionistic fuzzy decision tree. it is good.
  • the intuitionistic fuzzy random tree uses the Sigmoid function to fuzz the hard decision of the traditional decision tree, omits the complex fuzzy semantic process, and uses the intuitionistic fuzzy point operator to extend the fuzzy membership degree to the intuitionistic fuzzy membership degree and extract useful information. Improve Robustness.
  • the third embodiment of the target tracking method based on the intuitionistic fuzzy random forest is based on the first embodiment of the target tracking method based on the intuitionistic fuzzy random forest in the present invention, which is successful for association or matching.
  • Reliable Target Update Intuition Fuzzy Random Forest includes:
  • a reliable target with successful association or matching success is added to the positive training sample set as a new positive training sample in the current video frame, and several image blocks are selected as negative training samples within a specified range around the positive training sample.
  • the training sample set and the negative training sample constitute a training sample set.
  • the positive training sample set in this embodiment may include all corresponding image blocks in the current and previous video frames of the reliable target, and may also limit the number of positive training samples in the positive training sample set to be less than or equal to a specified threshold to save storage resources.
  • S42 Randomly sampling a plurality of samples from the training sample set to obtain a subset of the training samples.
  • S43 Generate an intuitionistic fuzzy decision tree by using the training sample subset.
  • step S42 If yes, the process ends, and the generated T intuitionistic fuzzy decision trees constitute a new intuitionistic fuzzy random forest; if not, return to step S42 to continue the loop.
  • This embodiment can be combined with the second embodiment of the target tracking method based on the intuitionistic fuzzy random forest of the present invention.
  • Step S43 specifically includes:
  • S431 Initializing the intuitionistic fuzzy membership degree of the training sample belonging to the root node in the training sample subset.
  • the intuitionistic fuzzy membership of the positive training sample belonging to the root node is 1/n 1
  • the total number n 0 of the negative training samples in the training sample subset is In the case of a positive integer, the intuitionistic fuzzy membership of the negative training sample belonging to the root node is 1/n 0 .
  • S432 Perform feature selection criterion training on the training samples reaching the current node.
  • the initial current node is the root node.
  • the optimal one-dimensional feature of the current node and the operator number of the optimal one-dimensional feature and the value of the feature threshold are confirmed.
  • the optimal one-dimensional feature belongs to the high-dimensional feature vector of the training sample.
  • Stop conditions can include:
  • step S434 If any one of the above three conditions is satisfied, then the process goes to step S434, and if none of them is satisfied, the process goes to step S435.
  • S435 Splitting the current node by using the optimal one-dimensional feature to generate two branch nodes of the next layer.
  • the process of generating an intuitionistic fuzzy decision tree is a process of recursively constructing a binary tree starting from the root node and maximizing the gain of intuitionistic fuzzy information as a feature selection criterion.
  • step S432 specifically includes:
  • S410 randomly select a one-dimensional feature from the high-dimensional feature vector of the training sample.
  • S420 Select one of the candidate feature thresholds, calculate the intuitionistic fuzzy information gain when the number of operators takes different values under the selected one-dimensional feature and the feature threshold, and record the selected one-dimensional feature and feature threshold. The value, the maximum intuitionistic fuzzy information gain, and the value of the corresponding operator number.
  • the candidate feature threshold may include a median value of two adjacent values obtained by sorting the values of the selected one-dimensional features of the training sample, and n training samples may obtain n- 1 median.
  • the candidate feature threshold may also include an average of the values of the selected one-dimensional features of all training samples. Of course, it can also be a combination of the above two.
  • the intuitionistic fuzzy information gain ⁇ H is defined as:
  • the training samples can pass through each output path of each branch node to each node in the intuitionistic fuzzy decision tree, so X is the initial training sample subset.
  • H(X) is the intuitionistic fuzzy entropy of set X, defined as:
  • ⁇ ( ⁇ ) is the Dirac function
  • c j is the category label of the training sample
  • the definition and calculation of the intuitionistic fuzzy membership degree h % (x j ) of the training sample belonging to the current node can be referred to the equations (4)-(9). It should be noted that the samples in the equations (4)-(9) at this time x is the training sample in the set X, and the sample features, feature thresholds, and number of operators used are belonging to the branch nodes before reaching the current node.
  • H l (X) is the intuitionistic fuzzy entropy of the set of training samples contained in the left branch of the current node output, defined as:
  • H r (X) is the intuitionistic fuzzy entropy of the set of training samples contained in the right branch of the current node output, defined as:
  • the intuitionistic fuzzy information gain ⁇ H is calculated separately when the number of operators is taken, and the largest intuitionistic fuzzy information gain ⁇ H is found therefrom for recording.
  • step S430 Perform a previous step for each of the candidate feature thresholds (ie, step S420), and find and save the one with the largest gain of the intuitionistic fuzzy information in all the records.
  • the one-dimensional feature included in the record is the optimal one-dimensional feature, the value of the feature threshold and the operator
  • the value of the number of times is the number of operators and the value of the feature threshold of the optimal one-dimensional feature.
  • Step S43 specifically includes:
  • S436 Initializing the intuitionistic fuzzy membership of the training sample subordinate to the root node in the training sample subset.
  • S437 Determine whether the current node meets the stop condition.
  • the initial current node is the root node.
  • step S438 If any of the above three conditions is satisfied, the process proceeds to step S438, and if none of them is satisfied, the process proceeds to step S439.
  • S439 Perform feature selection criterion training on the training samples reaching the current node, and use the optimal one-dimensional feature to split the current node to generate two branch nodes of the next layer.
  • the branch node is returned as the current node to step S437 to continue execution until all the current nodes become leaf nodes, and no branch nodes are generated.
  • the difference between this embodiment and the fourth embodiment of the object tracking method based on the intuitionistic fuzzy random forest of the present invention is that the step of determining whether the current node satisfies the stop condition and the step of performing the feature selection criterion training for the training sample reaching the current node are different.
  • the fourth embodiment of the target tracking method based on the intuitionistic fuzzy random forest of the present invention and details are not described herein again.
  • the sixth embodiment of the present invention is based on the first embodiment of the target tracking method based on the intuitionistic fuzzy random forest, and the step S2 includes:
  • the similarity measure includes a spatial distance feature similarity measure and an appearance feature similarity measure.
  • the spatial distance feature is one of the features that can more effectively match the observation and prediction results of the target.
  • the spatial distance feature similarity measure ⁇ 1 between the observation d and the prediction result o is defined as:
  • the appearance feature similarity measure ⁇ 2 between the observation d and the prediction result o is defined as:
  • s( ⁇ ) is a normalized correlation measure between the observation d and the target template e i , defined as:
  • d(x, y) is the gray value of the observation result d at the coordinates (x, y)
  • e i (x, y) is the gray value of the target template e i at the coordinates (x, y)
  • the observation d is also whitened and scaled to h ⁇ w.
  • the value of s ranges from [0,1].
  • the associated cost between the observation d and the predicted result o is defined as:
  • ⁇ ij is the correlation cost between the observation d i defined by equation (17) and the prediction result o j
  • the correlation matrix A 0 which minimizes the total associated cost of the observation result and the prediction result is the optimal correlation matrix.
  • the Hungarian algorithm can be used to solve the correlation results.
  • the seventh embodiment of the object tracking method based on the intuitionistic fuzzy random forest is based on the sixth embodiment of the target tracking method based on the intuitionistic fuzzy random forest of the present invention.
  • the method further includes:
  • the associated/matching object image block with the reliable target current frame association success or matching success is whitened and scaled to h ⁇ w and then added to the target template set of the reliable target. If the number of target templates in the target template set before joining is equal to the seventh threshold, the earliest added target template in the target template set is deleted.
  • step S4 may be independent of each other or may be performed simultaneously.
  • the first embodiment of the object tracking apparatus based on the intuitionistic fuzzy random forest of the present invention comprises: a processor 110 and a camera 120.
  • the camera 120 can be a local camera, the processor 110 is connected to the camera 120 via a bus; the camera 120 can also be a remote camera, and the processor 110 is connected to the camera 120 via a local area network or the Internet.
  • the processor 110 controls the operation of the target tracking device based on the intuitionistic fuzzy random forest, and the processor 110 may also be referred to as a CPU (Central Processing Unit).
  • Processor 110 may be an integrated circuit chip with signal processing capabilities.
  • the processor 110 can also be a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, and discrete hardware components.
  • the general purpose processor may be a microprocessor or the processor or any conventional processor or the like.
  • the object tracking device based on the intuitionistic fuzzy random forest may further include a memory (not shown) for storing instructions and data necessary for the operation of the processor 110, and may also store video data captured by the transmitter 120.
  • the processor 110 is configured to perform motion detection on the current video frame acquired from the camera 120, detect the obtained possible moving object as an observation result, and associate the observation result with the prediction result of the target, wherein the prediction result is at least using the previous video frame.
  • the target trajectory is predicted, and the target includes a reliable target and a temporary target; the trajectory management is performed on the unrelated observation result and the predicted result, which includes online tracking of the uncorrelated reliable target to obtain the candidate result,
  • the candidate results are matched by using the intuitionistic fuzzy random forest of the unassociated reliable target; the trajectory of the target of the current frame is obtained by using the correlation result and the matching result, including the prediction that the successful target with successful matching is predicted by the candidate result of the matching success
  • the result is filtered and updated to obtain the trajectory; the trajectory of the target of the current frame is used for prediction, and the intuitionistic fuzzy random forest is updated for the reliable target with successful association or matching success.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种基于直觉模糊随机森林的目标跟踪方法，包括：对当前视频帧进行运动检测，检测得到的可能运动对象作为观测结果；对观测结果和目标的预测结果进行关联，目标包括可靠目标及临时目标；对未被关联的观测结果和预测结果进行轨迹管理，其中包括对未被关联的可靠目标的预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配；利用关联结果和匹配结果获取当前帧的目标的轨迹，利用当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的可靠目标更新直觉模糊随机森林。本发明还公开了一种基于直觉模糊随机森林的目标跟踪装置。

Description

基于直觉模糊随机森林的目标跟踪方法及装置 【技术领域】
本发明涉及目标跟踪领域，特别是涉及一种基于直觉模糊随机森林的目标跟踪方法及装置。
【背景技术】
在线目标跟踪是计算机视觉中的一个热点研究课题，其对于动作识别、行为分析、场景理解等高层次的视觉研究具有重要意义，并且在视频监控、智能机器人、人机交互等领域有着广泛的应用前景。
在复杂场景下，由于目标自身形变、目标间相互遮挡或者背景静物对目标的遮挡等因素的影响，将难以避免的产生漏检。此时，漏检目标找不到与其关联的检测到的观测对象，无法通过数据关联为这些漏检目标的轨迹更新找到有效的信息，轨迹精度降低。
【发明内容】
为了至少部分解决以上问题，本发明提出了一种基于直觉模糊随机森林的目标跟踪方法。该基于直觉模糊随机森林的目标跟踪方法包括：对当前视频帧进行运动检测，检测得到的可能运动对象作为观测结果；对观测结果和目标的预测结果进行关联，其中预测结果是至少利用前一视频帧的目标的轨迹进行预测而得到的，目标包括可靠目标及临时目标；对未被关联的观测结果和预测结果进行轨迹管理，其中包括对未被关联的可靠目标的预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配；利用关联结果和匹配结果获取当前帧的目标的轨迹，其中包括对匹配成功的可靠目标利用其匹配成功的候选结果对其预测结果进行滤波更新以获取轨迹；利用当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的可靠目标更新直觉模糊随机森林。
为了至少部分解决以上问题，本发明提出了一种基于直觉模糊随机森林的目标跟踪装置。该基于直觉模糊随机森林的目标跟踪装置包括：处理器和摄像机，处理器连接摄像机；处理器用于对从摄像机获取的当前视频帧进行运动检 测，检测得到的可能运动对象作为观测结果；对观测结果和目标的预测结果进行关联，其中预测结果是至少利用前一视频帧的目标的轨迹进行预测而得到的，目标包括可靠目标及临时目标；对未被关联的观测结果和预测结果进行轨迹管理，其中包括对未被关联的可靠目标的预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配；利用关联结果和匹配结果获取当前帧的目标的轨迹，其中包括对匹配成功的可靠目标利用其匹配成功的候选结果对其预测结果进行滤波更新以获取轨迹；利用当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的可靠目标更新直觉模糊随机森林。
本发明的有益效果是：通过对未被关联的可靠目标的预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配，如果匹配成功，则利用匹配成功的候选结果对该可靠目标的预测结果进行滤波更新以获取其轨迹，使得在发生漏检，目标找不到关联的观测对象的情况下，可以使用直觉模糊随机森林找出与其匹配的可用于其轨迹滤波更新的候选结果，从而提高目标轨迹的精度，改善目标跟踪的性能。
【附图说明】
图1是本发明基于直觉模糊随机森林的目标跟踪方法第一实施例的流程图；
图2是本发明基于直觉模糊随机森林的目标跟踪方法第二实施例一个例子中分支节点的硬判决函数和模糊判决函数的示意图；
图3是本发明基于直觉模糊随机森林的目标跟踪方法第二实施例一个例子中分支节点的模糊判决函数和直觉模糊判决函数的示意图；
图4是本发明基于直觉模糊随机森林的目标跟踪方法第三实施例的流程图；
图5是本发明基于直觉模糊随机森林的目标跟踪方法第四实施例的流程图；
图6是本发明基于直觉模糊随机森林的目标跟踪方法第四实施例中特征选择准则训练的流程图；
图7是本发明基于直觉模糊随机森林的目标跟踪方法第五实施例的流程图；
图8是本发明基于直觉模糊随机森林的目标跟踪方法第六实施例的流程图；
图9是本发明基于直觉模糊随机森林的目标跟踪方法第七实施例的流程图；
图10是本发明基于直觉模糊随机森林的目标跟踪装置第一实施例的结构示意图。
【具体实施方式】
下面结合附图和实施例对本发明进行详细说明。
如图1所示，本发明基于直觉模糊随机森林的目标跟踪方法第一实施例包括：
S1：对当前视频帧进行运动检测。
使用帧差法、光流法、背景减除法等运动检测算法对当前视频帧进行运动检测，以从中找出属于运动前景的像素，辅以中值滤波和简单的形态学处理，最终得到当前视频帧中的可能运动对象作为观测对象。一个观测对象是当前视频帧中的一个图像块，一般而言，观测对象的形状为矩形。
S2：对观测结果和目标的预测结果进行关联。
目标包括稳定跟踪的可靠目标及不稳定跟踪的临时目标。本步骤中的目标状态，即每个目标被标记为可靠目标还是临时目标，是由前一视频帧的轨迹管理决定的。临时目标包括在前一视频帧为未被关联且不是匹配成功的候选结果的观测结果建立的新的目标，以及连续关联成功的帧数小于或者等于第一帧数阈值且未被删除的目标。可靠目标包括连续关联成功的帧数大于第一帧数阈值且未被删除的目标。目标的预测结果是至少利用前一视频帧的目标的轨迹进行预测而得到的。
S3：对未被关联的观测结果和预测结果进行轨迹管理，其中包括对于未被关联的可靠目标，对其预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配。
具体而言，在可靠目标的预测结果位置及其周围指定范围内选择若干个图像块作为候选结果，图像块的大小一般与预测结果的大小一致，指定范围的大小及候选结果的数量一般由经验值决定。候选结果可以包括在指定范围内的未被关联的观测结果。相邻的候选结果可以彼此不重叠，也可以部分重叠。使用未被关联的可靠目标的直觉模糊随机森林作为分类器，该分类器的分类结果有可靠目标和非可靠目标两类。计算候选结果作为测试样本隶属于可靠目标类别的直觉模糊隶属度。若该直觉模糊隶属度大于第一阈值η1，且候选结果与可靠目标的预测结果的外观特征相似性度量大于第二阈值η2，则匹配成功，0.5＜η1＜1且0.5＜η2＜1。
当候选结果的数量大于一时，在分别为每个候选结果计算其作为测试样本隶属于可靠目标类别的直觉模糊隶属度之后，可以分别判断每个候选结果的直觉模糊隶属度是否大于第一阈值η1，与预测结果的外观特征相似性度量是否大于 第二阈值η2，如果有至少两个候选结果满足前述两个条件，则选择其中直觉模糊隶属度最大的一个(直觉模糊隶属度相同则选择外观特征相似性度量最大的一个)作为匹配成功的候选结果用于后续的目标状态和直觉模糊随机森林的更新；当然，也可以从候选结果中选择直觉模糊隶属度最大的一个的候选目标，然后判断选中的候选结果的直觉模糊隶属度是否大于第一阈值η1，与预测结果的外观特征相似性度量是否大于第二阈值η2，如果满足前述两个条件则匹配成功。
此外，根据关联结果和匹配结果对目标状态进行更新，包括目标的建立、删除和状态修改。具体包括：为未被关联且不是匹配成功的候选结果的观测结果建立新的临时目标；将连续关联成功的帧数大于第一帧数阈值λ1的临时目标变为可靠目标；删除连续关联不成功的帧数大于第二帧数阈值λ2的临时目标；删除连续关联不成功的帧数大于第三帧数阈值λ3，且匹配结果为匹配失败的可靠目标，匹配结果为匹配失败是指利用直觉模糊随机森林计算候选结果作为测试样本隶属于可靠目标类别的直觉模糊隶属度小于或者等于第六阈值η6，且满足0＜η6≤η1。其中λ1为大于1的正整数，λ2和λ3均为正整数，且满足λ3≥λ2≥1。
S4：利用关联结果和匹配结果获取当前帧的目标的轨迹，利用当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的可靠目标更新直觉模糊随机森林。
对匹配成功的可靠目标利用其匹配成功的候选结果对其预测结果进行滤波更新以获取轨迹。此外对关联成功的目标利用其关联的观测结果对其预测结果进行滤波更新以获取轨迹，对新的临时目标将对应的观测结果作为轨迹，对关联不成功且未被删除的临时目标以及关联不成功且匹配不成功且未被删除的可靠目标将其预测结果作为轨迹。
然后利用当前帧的目标的轨迹进行预测，得到的结果可以作为目标的预测结果用于下一帧的目标跟踪。在本发明一个实施例中，使用卡尔曼滤波器对当前帧的目标的轨迹进行预测以获取下一帧的目标的预测结果，卡尔曼滤波器也可以用于对预测结果和对应的观测结果/候选结果进行滤波以获取目标的轨迹。
利用关联成功或匹配成功的可靠目标对应的目标图像块为其更新直觉模糊随机森林。目标图像块可以不包括目标的轨迹信息，例如为关联成功的观测对象或者匹配成功的候选结果，此时更新直觉模糊随机森林的步骤与前述目标的轨迹的获取及预测的步骤的执行顺序并无限制。目标图像块也可以包括目标的轨迹信息，例如为可靠目标的轨迹所在位置的图像块，此时更新直觉模糊随机森林的步骤应在前述目标的轨迹的获取的步骤之后执行。
通过上述实施例的实施，对未被关联的可靠目标的预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配，如果匹配成功，则利用匹配成功的候选结果对该可靠目标的预测结果进行滤波更新以获取其轨迹，使得在发生漏检，目标找不到关联的观测对象的情况下，可以使用直觉模糊随机森林找出与其匹配的可用于其轨迹滤波更新的候选结果，从而提高目标轨迹的精度，改善目标跟踪的性能。
本发明基于直觉模糊随机森林的目标跟踪方法第二实施例，是在本发明基于直觉模糊随机森林的目标跟踪方法第一实施例的基础上，候选结果作为测试样本隶属于可靠目标类别的直觉模糊隶属度P(c＝m|w)为：
Figure PCTCN2016110478-appb-000001
其中c为测试样本的类别标签，m为可靠目标类别，w为测试样本，T为直觉模糊随机森林中直觉模糊决策树的个数，M为生成直觉模糊随机森林的训练样本的类别集合，φ1(c＝m|w)为利用第t个直觉模糊决策树计算得到的测试样本w隶属于可靠目标类别m的直觉模糊隶属度。
直觉模糊决策树对分支节点输出判决进行了直觉模糊化，使得同一个样本会以不同的直觉模糊隶属度经过分支节点的输出左分支以及输出右分支，最终到达多个叶子节点。因此，直觉模糊决策树的分类结果需要综合考虑多个叶子节点的信息。φt(c＝m|w)定义为：
Figure PCTCN2016110478-appb-000002
在第t个直觉模糊决策树中，Bt为测试样本w到达的所有叶子节点构成的集合，b为测试样本w到达的一个叶子节点，h％(w)为将叶子节点b作为当前节点时测试样本w隶属于当前节点的直觉模糊隶属度，
Figure PCTCN2016110478-appb-000003
为叶子节点b预测类别为m的置信度，定义为：
Figure PCTCN2016110478-appb-000004
其中xj为到达叶子节点b的训练样本，共有nb个，cj为训练样本xj的类别，δ(·)为狄拉克函数，h％(xj)为将叶子节点b作为当前节点时训练样本xj隶属于当前节点的直觉模糊隶属度。
训练样本和测试样本隶属于当前节点的直觉模糊隶属度的计算方式相同， 样本x隶属于当前节点的直觉模糊隶属度h％(x)为其隶属于到达当前节点所经过的所有分支节点的输出路径的直觉模糊隶属度的乘积，具体定义为：
Figure PCTCN2016110478-appb-000005
其中D为样本到达当前节点之前经过的所有分支节点的集合，d为集合中的一个分支节点，l表示分支节点的输出左分支，r表示分支节点的输出右分支，
Figure PCTCN2016110478-appb-000006
为样本到达当前节点所经过的隶属于分支节点d的输出路径的直觉模糊隶属度。
样本包括测试样本和训练样本，训练样本包括正训练样本和负训练样本，正训练样本是指为目标类别的训练样本，负训练样本是指为非目标类别的训练样本。若当前节点为根节点，则D为空，无法使用式(4)计算h％(x)。在这种情况下，当样本x为测试样本时，h％(x)＝1，当样本x为正训练样本且正训练样本的总个数为n1且n1为正整数时，h％(x)＝1/n1，当样本x为负训练样本且负训练样本的总个数为n0且n0为正整数时，h％(x)＝1/n0。
样本到达当前节点所经过的隶属于分支节点d的输出路径的直觉模糊隶属度
Figure PCTCN2016110478-appb-000007
定义为：
Figure PCTCN2016110478-appb-000008
其中
Figure PCTCN2016110478-appb-000009
为样本隶属于分支节点d的输出左分支的直觉模糊隶属度，
Figure PCTCN2016110478-appb-000010
为样本隶属于分支节点d的输出右分支的直觉模糊隶属度。根据式(4)计算时，应当根据样本经过的分支节点d的输出路径是左分支还是右分支，从式(5)中
Figure PCTCN2016110478-appb-000011
和
Figure PCTCN2016110478-appb-000012
的表达式中选择对应的一个代入。h(xd)为分支节点d的直觉模糊输出判决函数。
传统的二叉决策树的分支节点采用硬判决，其分支节点输出判决函数的定义为：
Figure PCTCN2016110478-appb-000013
其中，xd为分支节点d的样本x的特征值，τ为特征门限值。0对应该分支节点输出左分支，1对应该分支节点输出右分支。采用S型函数(即Sigmoid函数)对式(19)表示的传统的硬判决函数进行模糊化。
Figure PCTCN2016110478-appb-000014
其中xd为分支节点d的样本x的特征值，τ为特征门限值，θ为用于控制Sigmoid函数倾斜程度的常量参数，σ为特征值的标准差。
举例说明，xd的取值范围为[0，1]，τ为0.4，θ＝0.25时模糊化前后的硬判决函数和模糊判决函数如图2所示。图中的虚线表示式(19)所定义的硬判决函数，其输出在特征门限处发生跳变；实线表示式(7)所定义的模糊判决函数，其输出根据样本的特征值单调连续变化，且在特征门限处等于0.5。
然后采用直觉模糊点算子，进一步将基于Sigmoid函数的模糊判决函数推广到直觉模糊判决函数。
假设U是一个非空集合，集合U的直觉模糊集合(IFS(U))A的定义为：
A＝{＜u，μA(u)，vA(u)＞|u∈U}          (20)
其中μA：U→[0，1]，μA(u)表示集合U中元素u属于A的隶属度，vA：U→[0，1]，vA(u)表示集合U中元素u属于A的非隶属度，且对任意u有：
Figure PCTCN2016110478-appb-000015
集合U中元素u属于A的模糊直觉指数定义为：
πA(u)＝1-μA(u)-vA(u)              (22)
模糊直觉指数πA(u)表示元素u相对于直觉模糊集A的不确定信息。如果πA(u)的值很小，说明元素u属于A的隶属度值相对精确；如果πA(u)的值很大，则说明元素u属于A的隶属度值具有较大的不确定性。与模糊集合相比，直觉模糊集合能够体现隶属、非隶属和模糊直觉指数三方面的信息，从而有利于更好地处理不确定性的信息。
为了更好地利用模糊直觉指数中的信息，引入直觉模糊点算子。对于任意u∈U，令αu，βu∈[0，1]，且满足αu+βu≤1，直觉模糊点算子
Figure PCTCN2016110478-appb-000016
定义为：
Figure PCTCN2016110478-appb-000017
直觉模糊点算子
Figure PCTCN2016110478-appb-000018
将直觉模糊集A转化为带有如下模糊直觉指数的直觉模糊集：
Figure PCTCN2016110478-appb-000019
记
Figure PCTCN2016110478-appb-000020
则有：
Figure PCTCN2016110478-appb-000021
Figure PCTCN2016110478-appb-000022
以此类推，可得对于任意正整数n，若
Figure PCTCN2016110478-appb-000023
αu+βu≠0，则有：
Figure PCTCN2016110478-appb-000024
Figure PCTCN2016110478-appb-000025
若对某个u∈U，αu+βu＝0，即αu＝0且βu＝0，则有：
Figure PCTCN2016110478-appb-000026
从式(27)以及式(28)可以看出，直觉模糊点算子
Figure PCTCN2016110478-appb-000027
将模糊直觉指数πA(u)划分为：(1-αu-βu)nπA(u)，αuπA(u)(1-(1-αu-βu)n)/(αu+βu)和βuπA(u)(1-(1-αu-βu)n)/(αu+βu)共三部分，分别表示在原不确定信息中的未知、隶属及非隶属部分。
对于任意u∈U，由于αu，βu∈[0，1]，且满足αu+βu≤1，有：
Figure PCTCN2016110478-appb-000028
式(30)表明，直觉模糊点算子
Figure PCTCN2016110478-appb-000029
能够减小直觉模糊集A的模糊直觉指数。这说明通过直觉模糊点算子
Figure PCTCN2016110478-appb-000030
可以从元素u相对于直觉模糊集A的不确定信息中提取出新的信息，提高不确定信息的利用程度。
经过直觉模糊推广得到的分支节点d的直觉模糊输出判决函数h(xd)定义为：
Figure PCTCN2016110478-appb-000031
其中k为算子次数，为正整数，k的取值越大，消耗的计算量越大。为了计算方便，可以取k∈{1，2，3}。
式(6)中的π(z)为模糊直觉指数，根据Sugeno模糊补，定义为：
Figure PCTCN2016110478-appb-000032
其中λ为常量参数，0＜λ＜1，例如为0.8。当xd≥τ时，z＝g(xd)，当xd＜τ时， z＝1-g(xd)。
式(6)中的α为从模糊直觉指数中提取隶属信息的尺度因子，β为从模糊直觉指数中提取非隶属信息的尺度因子，定义为：
Figure PCTCN2016110478-appb-000033
β＝1-α-π(z)
式(6)的输出值表示样本隶属于分支节点输出右分支的直觉模糊隶属度。直觉模糊点算子
Figure PCTCN2016110478-appb-000034
能够从不确定信息中提取新的有用信息，式(6)就是在原模糊隶属度信息g(xd)中加入了从模糊直觉指数π(z)中提取的隶属信息，从而减少了原模糊隶属度信息的不确定性。由于当k＝0时，由式(6)可以得到h(xd)＝g(xd)，此时，直觉模糊判决退化为模糊判决，因此，可以认为式(6)是对式(7)的直觉模糊推广。
举例说明，当k＝1，特征门限值τ＝0.4，λ＝0.8时，分支节点直觉模糊输出判决函数的图形如图3所示。图中虚线表示式(7)所定义的模糊判决函数；实线表示式(6)所定义的直觉模糊判决函数。
式(6)中算子次数k和特征门限值τ的取值可以通过更新直觉模糊决策树过程中的特征选择准则训练确定，也可以通过其他方式例如取经验值确定。
传统的二叉决策树的分支节点采用硬判决，测试样本只能根据特征属性从左右两个分支中选择一个到达下一层节点，最终到达一个叶子节点，测试样本的类别由该到达的叶子节点的类别决定。这样的硬判决决策树对样本噪声的鲁棒性不强，当样本受到强噪声干扰时，其特征值将会发生较大变化，可能导致样本经过的分支发生变化，降低决策树的准确性。
现有技术中提出了模糊决策树，将模糊集理论应用于决策树的训练以及推理过程，利用模糊集理论的表示能力来提高传统决策树对于带噪声的数据以及不完整的数据的处理能力。尽管模糊决策树能够处理带有不确定性的特征值，但是其需要对样本特征进行模糊语义化处理，而在目标跟踪中采用的样本特征大多为数值型特征，并且特征维数较高，使得对样本特征的模糊语义化处理变得困难。
本实施例中的直觉模糊随机森林综合考虑其中每个直觉模糊决策树所计算得到的直觉模糊隶属度来得到测试样本的分类结果，直觉模糊随机森林的分类性能比单个直觉模糊决策树的分类性能好。其中的直觉模糊随机树采用Sigmoid函数对传统决策树的硬判决进行模糊化，省略复杂的模糊语义化过程，并且采用直觉模糊点算子将模糊隶属度推广到直觉模糊隶属度，提取有用信息，提高 鲁棒性。
如图4所示，本发明基于直觉模糊随机森林的目标跟踪方法第三实施例，是在本发明基于直觉模糊随机森林的目标跟踪方法第一实施例的基础上，为关联成功或匹配成功的可靠目标更新直觉模糊随机森林包括：
S41：更新训练样本集合。
将关联成功或匹配成功的可靠目标在当前视频帧中对应的目标图像块作为新的正训练样本加入正训练样本集中，在正训练样本周围指定范围内选择若干个图像块作为负训练样本，正训练样本集和负训练样本组成训练样本集合。本实施例中的正训练样本集中可以包括可靠目标在当前以及之前视频帧中所有对应的图像块，也可以限制正训练样本集中正训练样本的数量小于或者等于指定阈值以节省存储资源。
S42：从训练样本集合中有放回地随机采样若干个样本以获取训练样本子集。
这意味着训练样本子集中可能有重复出现的同一训练样本，可以避免过度拟合。
S43：利用训练样本子集生成直觉模糊决策树。
S44：判断生成的直觉模糊决策树的数量是否到达预设数量T。
若是，则结束流程，生成的T个直觉模糊决策树组成了新的直觉模糊随机森林；若否则返回步骤S42继续循环。
本实施例可以与本发明基于直觉模糊随机森林的目标跟踪方法第二实施例相结合。
如图5所示，本发明基于直觉模糊随机森林的目标跟踪方法第四实施例，是在本发明基于直觉模糊随机森林的目标跟踪方法第三实施例的基础上，步骤S43具体包括：
S431：初始化训练样本子集中训练样本隶属于根节点的直觉模糊隶属度。
训练样本子集中正训练样本的总个数n1为正整数时，正训练样本隶属于根节点的直觉模糊隶属度为1/n1，训练样本子集中负训练样本的总个数n0为正整数时，负训练样本隶属于根节点的直觉模糊隶属度为1/n0。
S432：对到达当前节点的训练样本进行特征选择准则训练。
初始的当前节点为根节点。
根据直觉模糊信息增益最大原则确认当前节点的最优一维特征及最优一维特征的算子次数和特征门限值的取值，最优一维特征属于训练样本的高维特征矢量。
S433：判断当前节点是否满足停止条件。
直觉模糊决策树的深度越深，其消耗的存储资源越多，需要的计算量也越大，因此，需要设计其生成过程的停止条件。停止条件可以包括：
1)到达当前节点某一类别的训练样本隶属于当前节点的直觉模糊隶属度的和占到达当前节点全部训练样本的直觉模糊隶属度的总和的比重大于第三阈值θr；
2)到达当前节点的训练样本隶属于当前节点的直觉模糊隶属度的总和小于第四阈值θl；
3)当前节点在直觉模糊决策树中的深度达到第五阈值θd。
若满足以上三个条件中的任意一个，则跳转到步骤S434，若均不满足，则跳转到步骤S435.
S434：将当前节点转化为叶子节点。
S435：使用最优一维特征将当前节点分裂生成下一层的两个分支节点。
然后将分支节点作为当前节点返回步骤S432继续执行，直至所有的当前节点都变为叶子节点，不再有分支节点生成。直觉模糊决策树的生成过程是从根节点开始，以直觉模糊信息增益最大化作为特征选择准则，递归地构建二叉树的过程。
其中如图6所示，步骤S432具体包括：
S410：从训练样本的高维特征矢量中随机选择一个一维特征。
S420：从候选特征门限值中选择一个，在选中的一维特征和特征门限值条件下计算算子次数取不同数值时的直觉模糊信息增益，记录选中的一维特征、特征门限值的取值、最大的直觉模糊信息增益以及对应的算子次数的取值。
在本发明一个实施例中，候选特征门限值可以包括对训练样本的选中的一维特征的取值进行排序后得到的相邻两个取值的中值，n个训练样本可以得到n-1个中值。候选特征门限值也可以包括所有训练样本的选中的一维特征的取值的平均值。当然也可以是以上两种的组合。
直觉模糊信息增益ΔH定义为：
Figure PCTCN2016110478-appb-000035
其中X＝{x1，x2，…，xn}为到达当前节点的训练样本的集合，n为集合X中训练样本的个数。一般来说训练样本能够经过每个分支节点的每条输出路径到达直 觉模糊决策树中的每一个节点，因此X即为初始化训练样本子集。
H(X)为集合X的直觉模糊熵，定义为：
Figure PCTCN2016110478-appb-000036
其中δ(·)为狄拉克函数，cj为训练样本的类别标签，mi为训练样本的类别，由于仅需要对目标和非目标进行区分，因此共有两类，i＝1，2。训练样本隶属于当前节点的直觉模糊隶属度h％(xj)的定义及计算方式可参照式(4)-(9)，需要注意的是此时式(4)-(9)中的样本x为集合X中的训练样本，使用的样本特征、特征门限值和算子次数是属于到达当前节点之前的分支节点的。
Hl(X)为当前节点输出左分支所包含的训练样本的集合的直觉模糊熵，定义为：
Figure PCTCN2016110478-appb-000037
Hr(X)为当前节点输出右分支所包含的训练样本的集合的直觉模糊熵，定义为：
Figure PCTCN2016110478-appb-000038
其中
Figure PCTCN2016110478-appb-000039
为样本隶属于当前节点的输出左分支的直觉模糊隶属度，
Figure PCTCN2016110478-appb-000040
为样本隶属于当前节点的输出右分支的直觉模糊隶属度，其计算方式与分支节点d的相同，可参考式(5)-(9)，需要注意的是此时式(5)-(9)中的样本x仍为集合X中的训练样本，但使用是当前节点的选中的一维特征和特征门限值，以及本次指定的算子次数。
在算子次数的取不同数值时分别计算其直觉模糊信息增益ΔH，从中找出最大的直觉模糊信息增益ΔH以记录。
S430：为候选特征门限值中的每一个执行前一步骤(即步骤S420)，找出并保存所有记录中直觉模糊信息增益最大的一条。
该条记录中包括的一维特征即为最优一维特征，特征门限值的取值和算子 次数的取值即为最优一维特征的算子次数和特征门限值的取值。
如图7所示，本发明基于直觉模糊随机森林的目标跟踪方法第五实施例，是在本发明基于直觉模糊随机森林的目标跟踪方法第三实施例的基础上，步骤S43具体包括：
S436：初始化训练样本子集中训练样本隶属于根节点的直觉模糊隶属度。
S437：判断当前节点是否满足停止条件。
初始的当前节点为根节点。
若满足以上三个条件中的任意一个，则跳转到步骤S438，若均不满足，则跳转到步骤S439。
S438：将当前节点转化为叶子节点。
S439：对到达当前节点的训练样本进行特征选择准则训练，使用最优一维特征将当前节点分裂生成下一层的两个分支节点。
然后将分支节点作为当前节点返回步骤S437继续执行，直至所有的当前节点都变为叶子节点，不再有分支节点生成。
本实施例与本发明基于直觉模糊随机森林的目标跟踪方法第四实施例的区别在于判断当前节点是否满足停止条件的步骤与对到达当前节点的训练样本进行特征选择准则训练的步骤的执行顺序不同，具体内容可参考本发明基于直觉模糊随机森林的目标跟踪方法第四实施例，在此不再赘述。
如图8所示，本发明基于直觉模糊随机森林的目标跟踪方法第六实施例，是在本发明基于直觉模糊随机森林的目标跟踪方法第一实施例的基础上，步骤S2包括：
S21：计算观测结果和预测结果之间的相似性度量。
相似性度量包括空间距离特征相似性度量以及外观特征相似性度量。
通常，目标在相邻帧图像之间的位置不会发生较大变化，因此，空间距离特征是能够较为有效地匹配目标的观测结果与预测结果的特征之一。观测结果d与预测结果o之间的空间距离特征相似性度量ψ1定义为：
Figure PCTCN2016110478-appb-000041
其中||·||2为二范数，(xo，yo)为预测结果o的中心坐标，(xd，yd)为观测结果d的中心坐标，ho为预测结果o的高度，
Figure PCTCN2016110478-appb-000042
为方差常量，可以取
Figure PCTCN2016110478-appb-000043
由于目标的外观可能会随时间发生变化，单一的固定目标模板无法形成对目标外观的准确描述，因此，采用目标模板集来表示目标外观。预测结果o对应的目标模板集为
Figure PCTCN2016110478-appb-000044
其中的目标模板ei，i＝1，…，n2为经过白化处理且大 小缩放至h×w的之前n2个视频帧中的关联/匹配对象图像块，n2为目标模板集中包括的目标模板的总数。为了存储和计算方便，对目标模板集中所包含的目标模板的数量进行限制，n2小于或等于第七阈值γ，可以取γ＝5。
观测结果d与预测结果o之间的外观特征相似性度量ψ2定义为：
Figure PCTCN2016110478-appb-000045
其中s(·)为观测结果d与目标模板ei之间的归一化相关性度量，定义为：
Figure PCTCN2016110478-appb-000046
其中d(x，y)为观测结果d在坐标(x，y)处的灰度值，ei(x，y)为目标模板ei在坐标(x，y)处的灰度值，并且观测结果d也经过白化处理且大小缩放至h×w。s的取值范围为[0，1]。
S22：利用相似性度量计算观测结果和预测结果之间的关联代价。
观测结果d与预测结果o之间的关联代价定义为：
ρo，d＝1-ψ1×ψ2         (17)
S23：利用关联代价计算观测结果和预测结果之间的最优关联矩阵作为关联结果。
所有观测结果组成的集合为D＝{d1，…，dp}，所有预测结果组成的集合为O＝{o1，…，oq}，观测结果和预测结果的总关联代价定义为：
Figure PCTCN2016110478-appb-000047
其中ρij为式(17)定义的观测结果di与预测结果oj之间的关联代价，A＝[aij]p×q为观测结果和预测结果之间的关联矩阵，关联矩阵中的任一元素aij∈{0，1}，当aij＝1时，表示观测结果di与预测结果oj关联成功。
由于一个观测仅能够与一个目标相关联，并且一个目标也仅能够与一个观测相关联，求解
Figure PCTCN2016110478-appb-000048
得到使得观测结果和预测结果的总关联代价最小的关联矩阵A0即为最优关联矩阵。可以使用匈牙利算法求解得到关联结果。
如图9所示，本发明基于直觉模糊随机森林的目标跟踪方法第七实施例，是在本发明基于直觉模糊随机森林的目标跟踪方法第六实施例的基础上，步骤S3之后进一步包括：
S5：为关联成功或匹配成功的可靠目标更新目标模板集。
将可靠目标当前帧关联成功或匹配成功的关联/匹配对象图像块经过白化处理且大小缩放至h×w之后加入该可靠目标的目标模板集中。若加入之前目标模板集中目标模板的数量等于第七阈值，则删除目标模板集中最早加入的目标模板。
本步骤与步骤S4的执行可以是相互独立的，也可以是同时进行的。
如图10所示，本发明基于直觉模糊随机森林的目标跟踪装置第一实施例包括：处理器110和摄像机120。摄像机120可以为本地摄像机，处理器110通过总线连接摄像机120；摄像机120也可以为远程摄像机，处理器110通过局域网或互联网连接摄像机120。
处理器110控制基于直觉模糊随机森林的目标跟踪装置的操作，处理器110还可以称为CPU(Central Processing Unit，中央处理单元)。处理器110可能是一种集成电路芯片，具有信号的处理能力。处理器110还可以是通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
基于直觉模糊随机森林的目标跟踪装置可以进一步包括存储器(图中未画出)，存储器用于存储处理器110工作所必需的指令及数据，也可以存储传输器120拍摄的视频数据。
处理器110用于对从摄像机120获取的当前视频帧进行运动检测，检测得到的可能运动对象作为观测结果；对观测结果和目标的预测结果进行关联，其中预测结果是至少利用前一视频帧的目标的轨迹进行预测而得到的，目标包括可靠目标及临时目标；对未被关联的观测结果和预测结果进行轨迹管理，其中包括对未被关联的可靠目标的预测结果进行在线跟踪获取候选结果，利用未被关联的可靠目标的直觉模糊随机森林对候选结果进行匹配；利用关联结果和匹配结果获取当前帧的目标的轨迹，其中包括对匹配成功的可靠目标利用其匹配成功的候选结果对其预测结果进行滤波更新以获取轨迹；利用当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的可靠目标更新直觉模糊随机森林。
本发明基于直觉模糊随机森林的目标跟踪装置包括的各部分的功能可参考本发明在线目标跟踪方法各对应实施例中的描述，在此不再赘述。
以上所述仅为本发明的实施方式，并非因此限制本发明的专利范围，凡是利用本发明说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本发明的专利保护范围内。

Claims (19)

  1. 一种基于直觉模糊随机森林的目标跟踪方法，其中，包括：
    对当前视频帧进行运动检测，检测得到的可能运动对象作为观测结果；
    对所述观测结果和目标的预测结果进行关联，其中所述预测结果是至少利用前一视频帧的目标的轨迹进行预测而得到的，所述目标包括可靠目标及临时目标；
    对未被关联的所述观测结果和所述预测结果进行轨迹管理，其中包括对未被关联的所述可靠目标的预测结果进行在线跟踪获取候选结果，利用所述未被关联的所述可靠目标的直觉模糊随机森林对所述候选结果进行匹配；
    利用关联结果和匹配结果获取当前帧的目标的轨迹，其中包括对匹配成功的所述可靠目标利用其匹配成功的所述候选结果对其预测结果进行滤波更新以获取所述轨迹；利用所述当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的所述可靠目标更新所述直觉模糊随机森林。
  2. 根据权利要求1所述的方法，其中，
    所述对未被关联的所述可靠目标的预测结果进行在线跟踪获取候选结果包括：
    在所述可靠目标的预测结果位置及其周围指定范围内选择若干个图像块作为所述候选结果，所述候选结果可以包括未被关联的所述观测结果；
    所述利用所述未被关联的所述可靠目标的直觉模糊随机森林对所述候选结果进行匹配包括：
    利用所述直觉模糊随机森林计算所述候选结果作为测试样本隶属于所述可靠目标类别的直觉模糊隶属度；
    若所述直觉模糊隶属度大于第一阈值，且所述候选结果与所述可靠目标的预测结果的外观特征相似性度量大于第二阈值，则匹配成功。
  3. 根据权利要求2所述的方法，其中，
    所述候选结果作为测试样本隶属于所述可靠目标类别的直觉模糊隶属度P(c＝m|w)为：
    Figure PCTCN2016110478-appb-100001
    其中c为所述测试样本的类别标签，m为所述可靠目标类别，w为所述测试样本，T为所述直觉模糊随机森林中直觉模糊决策树的个数，M为生成所述直觉模糊随机森林的训练样本的类别集合，φt(c＝m|w)为利用第t个所述直觉模 糊决策树计算得到的所述测试样本w隶属于所述可靠目标类别m的直觉模糊隶属度，定义为：
    Figure PCTCN2016110478-appb-100002
    在所述第t个所述直觉模糊决策树中，Bt为所述测试样本w到达的所有叶子节点构成的集合，b为所述测试样本w到达的一个叶子节点，h％(w)为将所述叶子节点b作为当前节点时所述测试样本w隶属于所述当前节点的直觉模糊隶属度，
    Figure PCTCN2016110478-appb-100003
    为所述叶子节点b预测类别为m的置信度，定义为：
    Figure PCTCN2016110478-appb-100004
    其中xj为到达所述叶子节点b的训练样本，共有nb个，cj为所述训练样本xj的类别，δ(·)为狄拉克函数，h％(xj)为将所述叶子节点b作为当前节点时所述训练样本xj隶属于所述当前节点的直觉模糊隶属度。
  4. 根据权利要求3所述的方法，其中，
    样本x隶属于当前节点的直觉模糊隶属度定义h％(x)定义为：
    Figure PCTCN2016110478-appb-100005
    所述样本包括所述测试样本和所述训练样本，所述训练样本包括正训练样本和负训练样本，若所述当前节点为根节点，当所述样本x为所述测试样本时，h％(x)＝1，当所述样本x为所述正训练样本且所述正训练样本的总个数为n1且n1为正整数时，h％(x)＝1/n1，当所述样本x为所述负训练样本且所述负训练样本的总个数为n0且n0为正整数时，h％(x)＝1/n0；
    其中D为所述样本到达所述当前节点之前经过的所有分支节点的集合，d为所述集合中的一个分支节点，l表示所述分支节点的输出左分支，r表示所述分支节点的输出右分支，
    Figure PCTCN2016110478-appb-100006
    为所述样本到达所述当前节点所经过的隶属于所述分支节点d的输出路径的直觉模糊隶属度，定义为：
    Figure PCTCN2016110478-appb-100007
    其中
    Figure PCTCN2016110478-appb-100008
    为所述样本隶属于所述分支节点d的输出左分支的直觉模糊隶属度，
    Figure PCTCN2016110478-appb-100009
    为所述样本隶属于所述分支节点d的输出右分支的直觉模糊隶属度，h(xd)为所述分支节点d的直觉模糊输出判决函数，定义为：
    Figure PCTCN2016110478-appb-100010
    其中xd为所述分支节点的所述样本的特征值，k为算子次数，为正整数；
    g(xd)为S型函数，定义为：
    Figure PCTCN2016110478-appb-100011
    其中τ为特征门限值，θ为用于控制所述S型函数倾斜程度的常量参数，σ为所述特征值的标准差；
    π(·)为模糊直觉指数，定义为：
    Figure PCTCN2016110478-appb-100012
    其中λ为常量参数，0＜λ＜1，当xd≥τ时，z＝g(xd)，当xd＜τ时，z＝1-g(xd)；
    α为从模糊直觉指数中提取隶属信息的尺度因子，β为从模糊直觉指数中提取非隶属信息的尺度因子，定义为：
    Figure PCTCN2016110478-appb-100013
    β＝1-α-π(z)
    其中所述算子次数k和所述特征门限值τ的取值通过更新所述直觉模糊决策树过程中的特征选择准则训练确定。
  5. 根据权利要求1所述的方法，其中，
    所述为关联成功或匹配成功的所述可靠目标更新所述直觉模糊随机森林包括：
    将所述关联成功或匹配成功的所述可靠目标在所述当前视频帧中对应的目标图像块作为新的正训练样本加入正训练样本集中，在所述正训练样本周围指定范围内选择若干个图像块作为负训练样本，利用所述正训练样本集和所述负训练样本组成的训练样本集合生成新的所述直觉模糊随机森林。
  6. 根据权利要求5所述的方法，其中，
    所述利用所述正训练样本和所述负训练样本组成的训练样本集合生成新的所述直觉模糊随机森林包括：
    从所述训练样本集合中有放回地随机采样若干个样本以获取训练样本子集；
    利用所述训练样本子集生成直觉模糊决策树；
    循环执行上述步骤以生成预设数量个所述直觉模糊决策树，以组成所述新的直觉模糊随机森林。
  7. 根据权利要求6所述的方法，其中，
    所述利用所述训练样本子集生成直觉模糊决策树包括：
    初始化所述训练样本子集中训练样本隶属于根节点的直觉模糊隶属度，其中，所述训练样本子集中正训练样本的总个数n1为正整数时，所述正训练样本隶属于根节点的直觉模糊隶属度为1/n1，所述训练样本子集中负训练样本的总个数n0为正整数时，所述正训练样本隶属于根节点的直觉模糊隶属度为1/n0；
    对到达当前节点的所述训练样本进行特征选择准则训练，根据直觉模糊信息增益最大原则确认当前节点的最优一维特征及所述最优一维特征的算子次数和特征门限值的取值，其中所述最优一维特征属于所述训练样本的高维特征矢量，然后判断所述当前节点是否满足停止条件，若满足，则将所述当前节点转化为叶子节点，若不满足，则使用所述最优一维特征将所述当前节点分裂生成下一层的两个分支节点；或判断所述当前节点是否满足停止条件，若满足，则将所述当前节点转化为叶子节点，若不满足，则对到达当前节点的所述训练样本进行所述特征选择准则训练，然后使用所述最优一维特征将所述当前节点分裂生成下一层的两个分支节点；
    将所述分支节点作为当前节点返回前一步骤继续执行。
  8. 根据权利要求7所述的方法，其中，
    所述对到达当前节点的所述训练样本进行特征选择准则训练包括：
    从所述训练样本的高维特征矢量中随机选择一个一维特征；
    从候选特征门限值中选择一个，在选中的一维特征和特征门限值条件下计算所述算子次数取不同数值时的直觉模糊信息增益，记录所述选中的一维特征、所述特征门限值的取值、最大的所述直觉模糊信息增益以及对应的所述算子次数的取值；
    为所述候选特征门限值中的每一个执行前一步骤，找出并保存所有记录中所述直觉模糊信息增益最大的一条；
    重复执行上述步骤指定次数，在获取的所有保存的记录中找出所述直觉模糊信息增益最大的一条，其中包括的所述一维特征为所述最优一维特征，所述特征门限值的取值和所述算子次数的取值为所述最优一维特征的算子次数和特征门限值的取值。
  9. 根据权利要求8所述的方法，其中，
    所述直觉模糊信息增益ΔH定义为：
    Figure PCTCN2016110478-appb-100014
    其中X＝{x1，x2，...，xn}为到达所述当前节点的训练样本的集合，n为所述集合中所述训练样本的个数；
    H(X)为所述集合X的直觉模糊熵，定义为：
    Figure PCTCN2016110478-appb-100015
    其中δ(·)为狄拉克函数，cj为所述训练样本的类别标签，mi为所述训练样本的类别，由于仅需要对属于目标和不属于目标进行区分，因此共有两类，i＝1，2；
    训练样本隶属于所述当前节点的直觉模糊隶属度h％(·)定义为：
    Figure PCTCN2016110478-appb-100016
    其中D为所述样本到达所述当前节点之前经过的所有分支节点的集合，d为所述集合中的一个分支节点，
    Figure PCTCN2016110478-appb-100017
    为所述样本隶属于所述分支节点d的输出左分支的直觉模糊隶属度，
    Figure PCTCN2016110478-appb-100018
    为所述样本隶属于所述分支节点d的输出右分支的直觉模糊隶属度，定义为：
    Figure PCTCN2016110478-appb-100019
    其中h(xd)为所述分支节点d的直觉模糊输出判决函数，定义为：
    Figure PCTCN2016110478-appb-100020
    其中xd为所述分支节点的所述训练样本的特征值，k为所述算子次数，为正整数；
    g(xd)为S型函数，定义为：
    Figure PCTCN2016110478-appb-100021
    其中τ为所述特征门限值，θ为用于控制所述S型函数倾斜程度的常量参数，σ为所述特征值的标准差；
    π(·)为模糊直觉指数，定义为：
    Figure PCTCN2016110478-appb-100022
    其中λ为常量参数，0＜λ＜1，当xd≥τ时，z＝g(xd)，当xd＜τ时，z＝1-g(xd)；
    α为从模糊直觉指数中提取隶属信息的尺度因子，β为从模糊直觉指数中提取非隶属信息的尺度因子，定义为：
    Figure PCTCN2016110478-appb-100023
    β＝1-α-π(z)
    Hl(X)为所述当前节点输出左分支所包含的训练样本的集合的直觉模糊熵，定义为：
    Figure PCTCN2016110478-appb-100024
    Hr(X)为所述当前节点输出右分支所包含的训练样本的集合的直觉模糊熵，定义为：
    Figure PCTCN2016110478-appb-100025
    其中
    Figure PCTCN2016110478-appb-100026
    为所述样本隶属于所述当前节点的输出左分支的直觉模糊隶属度，
    Figure PCTCN2016110478-appb-100027
    为所述样本隶属于所述当前节点的输出右分支的直觉模糊隶属度，其计算方式与所述分支节点d的相同且使用的参数为所述选中的一维特征、特征门限值和算子次数。
  10. 根据权利要求8所述的方法，其中，所述候选特征门限值包括对所述训练样本的所述选中的一维特征的取值进行排序后得到的相邻两个所述取值的中值。
  11. 根据权利要求8所述的方法，其中，所述候选特征门限值包括所有所述训练样本的所述选中的一维特征的取值的平均值。
  12. 根据权利要求7所述的方法，其中，
    所述停止条件包括：
    到达所述当前节点某一类别的所述训练样本隶属于所述当前节点的直觉模 糊隶属度的和占到达所述当前节点全部所述训练样本的直觉模糊隶属度的总和的比重大于第三阈值；
    或到达所述当前节点的所述训练样本隶属于所述当前节点的直觉模糊隶属度的总和小于第四阈值；
    或所述当前节点在所述直觉模糊决策树中的深度达到第五阈值。
  13. 根据权利要求1-12中任一项所述的方法，其中，
    对未被关联的所述观测结果和所述预测结果进行轨迹管理进一步包括：
    为未被关联且不是匹配成功的候选结果的所述观测结果建立新的临时目标，将连续关联成功的帧数大于第一帧数阈值的所述临时目标变为可靠目标，删除连续关联不成功的帧数大于第二帧数阈值的所述临时目标，删除连续关联不成功的帧数大于第三帧数阈值，且所述匹配结果为匹配失败的所述可靠目标，其中所述匹配结果为匹配失败是指利用所述直觉模糊随机森林计算所述候选结果作为测试样本隶属于所述可靠目标类别的直觉模糊隶属度小于或者等于第六阈值。
  14. 根据权利要求13所述的方法，其中，
    所述利用关联结果和匹配结果获取当前帧的目标的轨迹进一步包括：
    对关联成功的所述目标利用其关联的所述观测结果对其预测结果进行滤波更新以获取所述轨迹，对所述新的临时目标将对应的所述观测结果作为所述轨迹，对关联不成功且未被删除的所述临时目标以及关联不成功且匹配不成功且未被删除的所述可靠目标将其预测结果作为所述轨迹。
  15. 根据权利要求1-12中任一项所述的方法，其中，
    所述对所述观测结果和目标的预测结果进行关联包括：
    计算所述观测结果和所述预测结果之间的相似性度量，所述相似性度量包括空间距离特征相似性度量以及外观特征相似性度量；
    利用所述相似性度量计算所述观测结果和所述预测结果之间的关联代价；
    利用所述关联代价计算所述观测结果和所述预测结果之间的最优关联矩阵作为关联结果，使得所述观测结果和所述预测结果的总关联代价最小。
  16. 根据权利要求15所述的方法，其中，包括：
    观测结果d与预测结果o之间的所述空间距离特征相似性度量ψ1定义为：
    Figure PCTCN2016110478-appb-100028
    其中||·||2为二范数，(xo，yo)为所述预测结果o的中心坐标，(xd，yd)为所述观测结果d的中心坐标，ho为所述预测结果o的高度，
    Figure PCTCN2016110478-appb-100029
    为方差常量；
    所述预测结果o对应的目标模板集为
    Figure PCTCN2016110478-appb-100030
    其中的目标模板ei，i＝1，...，n2为经过白化处理且大小缩放至h×w的之前n2个视频帧中的关联/匹配对象图像块，n2为所述目标模板集中包括的所述目标模板的总数且小于或等于第七阈值，所述观测结果d与所述预测结果o之间的所述外观特征相似性度量ψ2定义为：
    Figure PCTCN2016110478-appb-100031
    其中s(·)为所述观测结果d与所述目标模板ei之间的归一化相关性度量，定义为：
    Figure PCTCN2016110478-appb-100032
    其中d(x，y)为所述观测结果d在坐标(x，y)处的灰度值，ei(x，y)为所述目标模板ei在坐标(x，y)处的灰度值；
    所述观测结果d与所述预测结果o之间的关联代价定义为：
    ρo，d＝1-ψ1×ψ2              (17)
    所有所述观测结果组成的集合为D＝{d1，...，dp}，所有所述预测结果组成的集合为O＝{o1，...，oq}，所述观测结果和所述预测结果的总关联代价定义为：
    Figure PCTCN2016110478-appb-100033
    其中ρij为观测结果di与预测结果oj之间的关联代价，A＝[aij]p×q为所述观测结果和所述预测结果之间的关联矩阵，所述关联矩阵中的任一元素aij∈{0，1}，当aij＝1时，表示所述观测结果di与所述预测结果oj关联成功；所述最优关联矩阵为求解
    Figure PCTCN2016110478-appb-100034
    得到的关联矩阵A0。
  17. 根据权利要求16所述的方法，其中，进一步包括：
    对于关联成功或匹配成功的所述可靠目标，将其关联/匹配对象图像块经过白化处理且大小缩放至h×w之后加入所述可靠目标的所述目标模板集中，并且若加入之前所述目标模板集中所述目标模板的数量等于所述第七阈值，删除所述目标模板集中最早加入的所述目标模板。
  18. 根据权利要求1-12中任一项所述的方法，其中，
    所述利用所述当前帧的目标的轨迹进行预测包括：
    利用卡尔曼滤波器对所述当前帧的目标的轨迹进行预测以获取下一帧的目 标的预测结果。
  19. 一种基于直觉模糊随机森林的目标跟踪装置，其中，包括：处理器和摄像机，所述处理器连接所述摄像机；
    所述处理器用于对从所述摄像机获取的当前视频帧进行运动检测，检测得到的可能运动对象作为观测结果；对所述观测结果和目标的预测结果进行关联，其中所述预测结果是至少利用前一视频帧的目标的轨迹进行预测而得到的，所述目标包括可靠目标及临时目标；对未被关联的所述观测结果和所述预测结果进行轨迹管理，其中包括对未被关联的所述可靠目标的预测结果进行在线跟踪获取候选结果，利用所述未被关联的所述可靠目标的直觉模糊随机森林对所述候选结果进行匹配；利用关联结果和匹配结果获取当前帧的目标的轨迹，其中包括对匹配成功的所述可靠目标利用其匹配成功的所述候选结果对其预测结果进行滤波更新以获取所述轨迹；利用所述当前帧的目标的轨迹进行预测，并为关联成功或匹配成功的所述可靠目标更新所述直觉模糊随机森林。
PCT/CN2016/110478 2016-12-16 2016-12-16 基于直觉模糊随机森林的目标跟踪方法及装置 WO2018107492A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/110478 WO2018107492A1 (zh) 2016-12-16 2016-12-16 基于直觉模糊随机森林的目标跟踪方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/110478 WO2018107492A1 (zh) 2016-12-16 2016-12-16 基于直觉模糊随机森林的目标跟踪方法及装置

Publications (1)

Publication Number Publication Date
WO2018107492A1 true WO2018107492A1 (zh) 2018-06-21

Family

ID=62557858

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/110478 WO2018107492A1 (zh) 2016-12-16 2016-12-16 基于直觉模糊随机森林的目标跟踪方法及装置

Country Status (1)

Country Link
WO (1) WO2018107492A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110390400A (zh) * 2019-07-02 2019-10-29 北京三快在线科技有限公司 计算模型的特征生成方法、装置、电子设备和存储介质
CN111292346A (zh) * 2020-01-16 2020-06-16 安徽理工大学 一种噪声环境下浇铸箱体轮廓的检测方法
CN111815760A (zh) * 2020-06-23 2020-10-23 中国地质大学（武汉） 一种三维地层可钻性场空间建模方法
CN113177615A (zh) * 2021-06-01 2021-07-27 西北工业大学 一种基于证据森林的信息不确定条件目标意图识别方法
CN116718197A (zh) * 2023-08-09 2023-09-08 腾讯科技（深圳）有限公司 轨迹处理方法、装置、电子设备及存储介质
CN116973945A (zh) * 2023-09-21 2023-10-31 山东科技大学 基于智能终端gnss数据的干扰检测方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102881022A (zh) * 2012-07-20 2013-01-16 西安电子科技大学 基于在线学习的遮挡目标跟踪方法
CN102945554A (zh) * 2012-10-25 2013-02-27 西安电子科技大学 基于学习和加速鲁棒surf特征的目标跟踪方法
CN105894542A (zh) * 2016-04-26 2016-08-24 深圳大学 一种在线目标跟踪方法及装置
US20160328613A1 (en) * 2015-05-05 2016-11-10 Xerox Corporation Online domain adaptation for multi-object tracking

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102881022A (zh) * 2012-07-20 2013-01-16 西安电子科技大学 基于在线学习的遮挡目标跟踪方法
CN102945554A (zh) * 2012-10-25 2013-02-27 西安电子科技大学 基于学习和加速鲁棒surf特征的目标跟踪方法
US20160328613A1 (en) * 2015-05-05 2016-11-10 Xerox Corporation Online domain adaptation for multi-object tracking
CN105894542A (zh) * 2016-04-26 2016-08-24 深圳大学 一种在线目标跟踪方法及装置

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110390400A (zh) * 2019-07-02 2019-10-29 北京三快在线科技有限公司 计算模型的特征生成方法、装置、电子设备和存储介质
CN110390400B (zh) * 2019-07-02 2023-07-14 北京三快在线科技有限公司 计算模型的特征生成方法、装置、电子设备和存储介质
CN111292346A (zh) * 2020-01-16 2020-06-16 安徽理工大学 一种噪声环境下浇铸箱体轮廓的检测方法
CN111292346B (zh) * 2020-01-16 2023-05-12 安徽理工大学 一种噪声环境下浇铸箱体轮廓的检测方法
CN111815760A (zh) * 2020-06-23 2020-10-23 中国地质大学（武汉） 一种三维地层可钻性场空间建模方法
CN113177615A (zh) * 2021-06-01 2021-07-27 西北工业大学 一种基于证据森林的信息不确定条件目标意图识别方法
CN113177615B (zh) * 2021-06-01 2024-01-19 西北工业大学 一种基于证据森林的信息不确定条件目标意图识别方法
CN116718197A (zh) * 2023-08-09 2023-09-08 腾讯科技（深圳）有限公司 轨迹处理方法、装置、电子设备及存储介质
CN116718197B (zh) * 2023-08-09 2023-10-24 腾讯科技（深圳）有限公司 轨迹处理方法、装置、电子设备及存储介质
CN116973945A (zh) * 2023-09-21 2023-10-31 山东科技大学 基于智能终端gnss数据的干扰检测方法及系统
CN116973945B (zh) * 2023-09-21 2023-12-08 山东科技大学 基于智能终端gnss数据的干扰检测方法及系统

Similar Documents

Publication Publication Date Title
CN106846355B (zh) 基于提升直觉模糊树的目标跟踪方法及装置
WO2018107488A1 (zh) 基于提升直觉模糊树的目标跟踪方法及装置
WO2018107492A1 (zh) 基于直觉模糊随机森林的目标跟踪方法及装置
CN106846361B (zh) 基于直觉模糊随机森林的目标跟踪方法及装置
Yao et al. Deep discriminative cnn with temporal ensembling for ambiguously-labeled image classification
JP4767595B2 (ja) 対象物検出装置及びその学習装置
JP5558412B2 (ja) 識別器を特定のシーン内のオブジェクトを検出するように適応させるためのシステム及び方法
CN107145862B (zh) 一种基于霍夫森林的多特征匹配多目标跟踪方法
JP6598746B2 (ja) 他の物体の画像領域も考慮して物体を追跡する装置、プログラム及び方法
CN112070058A (zh) 人脸面部复合情感表情识别方法及系统
US11410327B2 (en) Location determination apparatus, location determination method and computer program
CN113283368B (zh) 一种模型训练方法、人脸属性分析方法、装置及介质
Bronskill et al. Memory efficient meta-learning with large images
Avola et al. A shape comparison reinforcement method based on feature extractors and f1-score
CN111444816A (zh) 一种基于Faster RCNN的多尺度密集行人检测方法
Akansha et al. Analytical review on video-based human activity recognition
Zhang et al. Transfer learning‐based online multiperson tracking with Gaussian process regression
Roy et al. Background subtraction using dominant directional pattern
CN112507924A (zh) 一种3d手势识别方法、装置及系统
CN113591607B (zh) 一种车站智能疫情防控系统及方法
Abayomi-Alli et al. Facial image quality assessment using an ensemble of pre-trained deep learning models (EFQnet)
Susan et al. Fuzzy C-means with non-extensive entropy regularization
González et al. Silhouette-based human action recognition with a multi-class support vector machine
Soni et al. Fine-grained uncertainty modeling in neural networks
Kavitha et al. An extreme learning machine and action recognition algorithm for generalized maximum clique problem in video event recognition

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16923658

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 16923658

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 07.10.2019)

122 Ep: pct application non-entry in european phase

Ref document number: 16923658

Country of ref document: EP

Kind code of ref document: A1